Minor improvements & docs added

.gitignore

@@ -13,5 +13,4 @@ main.todo
 sample.mp3
 *.spec
 
-# лютый завоз
 ui/assets/question_mark.png

env_cpu.yml

@@ -1,19 +0,0 @@
-name: notecast
-channels:
-  - pytorch
-  - conda-forge
-  - defaults
-dependencies:
-  - python=3.10
-  - pytorch::pytorch=2.0.1
-  - pytorch::torchvision=0.15.2
-  - pytorch::torchaudio=2.0.2
-  - transformers=4.30.0
-  - accelerate=0.20.0
-  - ffmpeg
-  - pyinstaller
-  - customtkinter
-  - isort
-  - mypy
-  - black
-  - tqdm

env_cuda.yml

@@ -1,15 +0,0 @@
-name: notecast
-channels:
-  - pytorch
-  - nvidia
-  - conda-forge
-  - bioconda
-dependencies:
-  - pytorch
-  - ffmpeg
-  - torchvision
-  - torchaudio
-  - transformers
-  - python=3.12
-  - customtkinter
-  - openai

env_mps.yml

@@ -1,19 +0,0 @@
-name: notecast
-channels:
-  - pytorch
-  - conda-forge
-  - defaults
-dependencies:
-  - python=3.10
-  - pytorch::pytorch=2.0.1
-  - pytorch::torchvision=0.15.2
-  - pytorch::torchaudio=2.0.2
-  - transformers=4.30.0
-  - accelerate=0.20.0
-  - ffmpeg
-  - pyinstaller
-  - customtkinter
-  - isort
-  - mypy
-  - black
-  - tqdm

transcription/audio_transcription.py

@@ -15,7 +15,54 @@ from ui.ui_log_handler import UILogHandler
 
 # TODO: implement transcription with shift
 class AudioTranscription:
-    model_name = "openai/whisper-large-v2"
+    """
+    Class for automatical audio transcription using Whisper.
+    
+    Provides audio file loading, resampling, conversion to mono,
+    splitting into chunks and batches, running the model, and compiling the final text.
+
+    Attributes: 
+        model_name (str): Whisper model name in HuggingFace.
+        filepath (str): Input filepath.
+        waveform (torch.Tensor): Audiosignal in tensor form.
+        sampling_rate (int): Input file's sampling frequency.
+        chunks (list): Audio's chunks list.
+        batches (list): Batches combined from chunks.
+        chunk_size (int): Chunk size in samples.
+        custom_chunk_length (int): Custom chunk length (in seconds).
+        custom_batch_length (int): Custom batch length (in chunks).
+        device (str): Inference device ("cuda", "cpu", "mps").
+        processor (WhisperProcessor | None): Tokenizator/preprocessor.
+        model (WhisperForConditionalGeneration | None): Whisper model.
+        logger (logging.Logger): Logger.
+        torch_dtype (torch.dtype): Data type for calculations (fp16/fp32).
+        language (str): Transcription language (default "ru").
+        all_transcription (list): List of strings with transcription.
+
+    Args:
+        filepath (str): Input filepath.
+        device_configuration (DeviceConfiguration): Device configuration 
+            (GPU/CPU/MPS, model, chunk length, batch etc.).
+        logger (logging.Logger): Logger.
+        language (str, optional): Transcription language. Default "ru".
+        
+    Example: 
+        >>> from transcription.device_configuration import DeviceConfiguration
+        >>> import logging
+        >>> config = DeviceConfiguration(device="cuda", model_name="openai/whisper-large-v3-turbo") # recheck this, not true i think
+        >>> logger = logging.getLogger("transcription")
+        >>> transcriber = AudioTranscription("audio.wav", config, logger, language="en")
+        >>> text = transcriber.transcribe_audio()
+        >>> print(text)
+        "This is a test audio transcription."
+    
+    Methods:
+        transcribe_audio() -> str:
+            Starts full transcription pipeline: model loading, 
+            file loading and preprocessing, splitting into chunks/batches, 
+            inference and model unloading. Returns the final transcription.
+    """
+    model_name = "openai/whisper-large-v3-turbo"
 
     filepath: str
     waveform: torch.Tensor
@@ -93,12 +140,14 @@ class AudioTranscription:
     def _load_file(self) -> None:
         self.logger.info(f"Loading file {self.filepath}")
         try:
-            self.waveform, self.sampling_rate = torchaudio.load(
-                self.filepath, format=self.file_format, backend="ffmpeg"
-            )
+            # self.waveform, self.sampling_rate = torchaudio.load(
+            #     self.filepath, format=self.file_format, backend="ffmpeg"
+            # )
+            self.waveform, self.sampling_rate = torchaudio.load(self.filepath)
             self.logger.info(f"Successfully loaded file {self.filepath}.")
         except Exception as e:
             self.logger.error(f"Unable to load file {self.filepath}: {e}")
+            # raise RuntimeError(f"Failed to load file {self.filepath}: {e}")
 
     def _resample(self) -> None:
         self.waveform = torchaudio.functional.resample(

transcription/device_configuration.py

@@ -17,6 +17,7 @@ class DeviceConfiguration:
             - "openai/whisper-medium"
             - "openai/whisper-large"
             - "openai/whisper-large-v2"
+            - "openai/whisper-large-v3-turbo"
 
         batch_size (int): Chunks in one batch. Selected for VRAM.
 
@@ -24,8 +25,8 @@ class DeviceConfiguration:
 
         data_type (str): custom data type of model. Variants:
             - torch.float16 - for GPUs
-            - torch.float32 - for CPU / weak GPU
-            - torch.bfloat16 - for GPUs which has BF16 support
+            - torch.float32 - for CPU
+            - torch.bfloat16 - for GPUs which has BF16 support (usually RTX 40XX+)
     """
 
     device: str = "cuda"

ui/ui.py

@@ -94,6 +94,7 @@ class TranscriberApp(ctk.CTk):
         self.device_var = tk.StringVar(value=device_opts[0])
         self.device_opts = device_opts
 
+        # input & output file variables
         self.input_file_var = tk.StringVar()
         self.output_file_var = tk.StringVar()
 
@@ -218,7 +219,6 @@ class TranscriberApp(ctk.CTk):
             tooltip="Choose model for speed recognition",
             variable=self.model_var,
             values=[
-                # maybe delete this option?
                 "openai/whisper-large-v3-turbo",
                 "openai/whisper-large-v2",
                 "openai/whisper-large",
@@ -257,7 +257,7 @@ class TranscriberApp(ctk.CTk):
             text="Chunk length (s):",
             tooltip="Maximum length of processing audio fragment",
             variable=self.chunk_var,
-            values=["30", "25", "20", "15", "10", "5"],
+            values=["30", "24", "20", "14", "10", "6"],
         )
 
         # third row